A Metric Learning Based Solution for Non-Stationary Acoustic Source Classification

Abstract

In this work, a metric learning-based approach is proposed for non-stationary acoustic source classification. A classic time-frequency representation of acoustic signals is adopted as input of a convolutional neural network in order to generate embedded features of reduced size. The embedding generation is optimized on similarity constraints in order to maximize intra-class and minimize inter-class distances. Eight sources with different degrees of non-stationarity are selected for the acoustic source classification task. Experiments demonstrated that the proposed solution outperforms the baseline systems for all individual acoustic sources, leading to an increment in the average balanced accuracy of more than twenty percentage points.

Non-Stationarity Objective Assessment for Acoustic Source Classification

“…It is important to notice that a balanced accuracy reduction is observed specially for T h /T = 0.4, i.e, considering nine INS features. This is can be partially explained by the fact that acoustic sources usually present a reduction on its non-stationary behavior for higher T h /T values [10] [4] [5]. Therefore the last incorporated feature does not hold the same discrimination power over features of reduced scale.…”

“…In this work, the Index of Non-Stationarity (INS) [9] is proposed as a complementary feature to improve individual non-stationary acoustic source recognition. It is known that different acoustic sources present distinct degrees of nonstationarity [4] [5] [10]. The idea is to incorporate the nonstationary pattern, intrinsic to every source, as an acoustic feature for classification systems.…”

Improving Non-Stationary Acoustic Source Classification with Metric Learning

“…In this work, the metric learning strategy is explored to improve individual non-stationary acoustic source classification. The idea is extended from [12] to overcome the statistical differences that arise from the non-stationary behavior by learning an embedding generator network optimal strategy that minimizes intra-class and maximizes inter-class distances [13]. A deep convolutional neural network (CNN) is adopted to extract embedded features of reduced size from time-frequency representations of acoustic signals.…”